Multiscale and Multi-Timestep Switching of Multiple Machine Learning Force Fields for Artificial Intelligence-Driven Materials Simulations

Molecular dynamics (MD) is a crucial technique in materials science, though its application to large systems and long timescales remains constrained by the prohibitive computational cost of high-accuracy simulations. To address this issue, we propose a multiscale MD approach that switches between two deep potential (DP) models, a type of machine learning force field (MLFF), with different precisions and speeds to optimally balance efficiency and accuracy. A high-precision DP model with a 6 Å cutoff and a faster, lower-precision DP model with a 4 Å cutoff are applied in a 1:3 ratio during integration.Evaluated on a TiO₂ crystal system, the proposed method preserved structural accuracy with Pearson correlation coefficients ≥0.995 for radial distribution functions (RDFs), while delivering 1.32× speedup compared to the high-precision baseline. Similarly, in a liquid polyethylene glycol (PEG) system, the method maintained RDF correlations ≥0.997 with a 1.27× speedup. Furthermore, when combining the switching scheme with network-size reduction (model compression) and mixed-precision (fp16) inference, RDF correlations were maintained at ≥0.99 while achieving a 3.95× speedup. These results demonstrate that the proposed method can substantially accelerate MD simulations without compromising accuracy, thereby offering a robust approach for artificial intelligence (AI)-assisted material design and large-scale simulations.